Machines in the Ghost Invited Position Paper for ENF 2007 Emulating the Mind 1st international Engineering and Neuro-Psychoanalysis Forum Vienna July 2007 http://www.indin2007.org/enf/ DRAFT: August 30, 2007(MAY CHANGE) Aaron Sloman http://www.cs.bham.ac.uk/∼axs/ School of Computer Science, University of Birmingham, UK Index Terms— architecture, artiﬁcial-intelligence, autonomy, is the common assumption that we already know what needsbehaviour, design-based, emotion, evolution, ghost in machine, to be explained and modelled, leaving only the probleminformation-processing, language, machine, mind, robot, philos- of ﬁnding good theories and designs for working systems.ophy, psychotherapy, virtual machine. Alas, what needs to be explained is itself a topic still Abstract— This paper summarises ideas I have been workingon over the last 35 years or so, about relations between the requiring much research, as acknowledged by the Researchstudy of natural minds and the design of artiﬁcial minds, and Roadmap project in the euCognition network (http://www.the requirements for both sorts of minds. The key idea is eucognition.org).1 .that natural minds are information-processing virtual machines A less obvious, but even more serious impediment toproduced by evolution. What sort of information-processing progress is the common assumption that our ordinary languagemachine a human mind is requires much detailed investigationof the many kinds of things minds can do. At present, it is is sufﬁcient for describing everything that needs to benot clear whether producing artiﬁcial minds with similar powers explained, leading to over-reliance on common-sense concepts.will require new kinds of computing machinery or merely much When scientists and engineers discuss what needs to befaster and bigger computers than we have now. Some things once explained, or modelled, and when they report experimentalthought hard to implement in artiﬁcial minds, such as affective observations, or propose explanatory theories, they often usestates and processes, including emotions, can be construed asaspects of the control mechanisms of minds. This view of mind concepts (such as ‘conscious’, ’unconscious’, ‘experience’,is largely compatible in principle with psychoanalytic theory, ‘emotion’, ‘learn’, ‘motive’, ‘memory’) that evolved not forthough some details are very different. The therapeutic aspect the purposes of science, but for use in informal discourseof psychoanalysis is analogous to run-time debugging of a among people engaged in every day social interaction, likevirtual machine. In order to do psychotherapy well we need to this:understand the architecture of the machine well enough to knowwhat sorts of bugs can develop and which ones can be removed, • What does the infant/child/adult/chimp/crow (etc)or have their impact reduced, and how. Otherwise treatment will perceive/understand/learn/intend (etc)?be a hit-and-miss affair. • What is he/she/it conscious of? • What does he/she/it experience/enjoy/desire? • What is he/she/it attending to? I. T HE DESIGN - BASED APPROACH TO STUDYING MINDS • What sort of emotion is he/she/it having now?M Y PRIMARY goal is to understand natural minds of all kinds, not to make smart machines. I ammore a philosopher than an engineer – though I have ﬁrst- These everyday usages may be ﬁne for everyday chats, gossip, consulting rooms, or even law courts and medical reports, but it does not follow that the concepts used (e.g.hand experience of software engineering. Deep scientiﬁc and ‘conscious’, ‘experience’, ‘desire’, ‘attend’, ‘emotion’) are anyphilosophical understanding of natural minds requires us to more adequate as concepts to be used in scientiﬁc theories thandescribe minds with sufﬁcient precision to enable our theories the everyday concepts of ‘mud’, ‘cloudy’, ‘hot’, ‘vegetable’,to be the basis for designs for working artiﬁcial minds like ‘reddish’ or ‘smelly’ are useful in theories about physics orours. Such designs can be compared with psychoanalytic chemistry.and other theories about how minds work. If a theory abouthow natural minds work is to be taken seriously, it should From vernacular to deep conceptsbe capable of providing the basis for the design of artiﬁcal The history of the physical sciences shows that as weworking minds. understand more about the architecture of matter and the Part of the problem is describing what needs to be variety of states, processes, and causal interactions that lieexplained, and deciding which concepts to use in formulating hidden from ordinary observation, the more we have toexplanatory theories. A very serious impediment to progress construct new systems of concepts and theories that make use This work was partly supported by the EU CoSy project. 1 Available online: ‘What’s a Research Roadmap For? Why do we need one? A. Sloman is at the School of Computer Science, University of Birming- How can we produce one?’ http://www.cs.bham.ac.uk/research/ham, UK, http://www.cs.bham.ac.uk/˜axs/ projects/cosy/papers/#pr0701
of them (and thereby help to deﬁne the concepts), in order between theoretical concepts and observable phenomena isto obtain deep explanations of what we can observe and new referred to as ‘symbol attachment’ in  and ‘symbolmore reliable and more precise predictions, regarding a wider tethering’ in . Once this possibility is understood, the needrange of phenomena. In particular, in physics, chemistry and for so-called ‘symbol grounding’ (deriving all concepts frombiology, new theories have taught us that however useful our experience) presented in , evaporates. (The impossibility ofordinary concepts are for ordinary everyday social interactions, ‘concept empiricism’ was demonstrated by Kant  over 200they are often grossly inadequate: (a) for distinguishing years ago.)all the phenomena that need to be distinguished, (b) for The ontology used by an individual or a community (i.e.formulating precise and reliable predictions, and (c) for the set of concepts used to describe things in the world) candescribing conditions (often unobservable conditions) that are be extended in two ways, either by deﬁnitional abbreviationthe basis for reliable predictions. or substantively. A deﬁnitional abbreviation merely introduces So it may be unwise to go on using the old concepts a new symbol as a short-hand for what could be expressedof folk psychology when describing laboratory experiments previously, whereas substantive ontology extensions introduceor ﬁeld observations, when formulating descriptions of what new concepts that cannot be deﬁned in terms of pre-existingneeds to be explained, and when formulating supposedly concepts. Such concepts are implicitly deﬁned mainly by theirexplanatory theories. But choosing alternatives to ordinary role in explanatory theories, as explained above. So substantivelanguage needs great care. Recently attempts have been made ontology extension always requires theory construction, as hasto give these concepts scientiﬁc status by using modern happened many times in the history of science and culture.technology to identify precise brain mechanisms, brain states,brain processes that correlate with the states and processesdescribed in ordinary language. Compare trying to identify the II. V IRTUAL AND PHYSICAL MACHINESprecise location in a particular country where some national We can do for the study of mind and brain what wascharacteristic or process such as religious bigotry, scientiﬁc previously done for the study of physical matter and biologicalignorance, or economic inﬂation is located. processes of evolution and development, if we understand thatThe ‘design-based’ approach minds and brains are not just matter-manipulating, or energy-Is there any alternative to going on using pre-scientiﬁc contexts manipulating machines, but information-processing machines.in our descriptions and theories? Yes, but it is not a simple Minds are information processing virtual machines whilealternative. We need to adopt what Dennett  calls ‘the brains are physical machines, in which those virtual machinesdesign stance’, which involves constructing theories about how are implemented or some would say realised.minds and brains actually work, which goes far beyond what In computing systems we also have virtual machines,we either experience of their working in ourselves or observe such as running operating systems, ﬁrewalls, email systems,in others. (My own early attempts at doing this 30 years ago spelling correctors, conﬂict resolution mechanisms, ﬁleare presented in .) Moreover, our theories must account for optimisation mechanisms, all of which run on, i.e. arethe existence of many kinds of minds, with different designs. implemented in, the underlying physical hardware. There are In particular we can’t just start deﬁning precise new no simple mappings between the components of the virtualexplanatory concepts in terms of precise measurements and and the physical machines. There are often several layersobservations. Thinking that deﬁnitions come before theories is of implementation between high level virtual machines anda common mistake. In the more advanced sciences, concepts physical machines. This provides enormous ﬂexibility for re-are used (e.g. ‘electron’, ‘charge’, ‘atomic weight’, ‘valency’, use of hardware, both in different systems using the same‘oxidation’, ‘gene’, etc.) that cannot be deﬁned except by hardware, and in one system performing different functionstheir role in the theories that employ them. The structural at different times. It seems that evolution discovered therelations within the theory partially deﬁne the concepts. That power of virtual machines before we did and produced brainsis because a theory is a formal system, and, as is familiar from implementing hierarchies of mental (and indirectly social)mathematics, and made more precise in the work of Tarski, virtual machines.the structure of a formal system determines which things The various components of a complex virtual machineare possible models of that system. Usually there are many (such as an operating system distributed over a networkpossible models, but the set of possible models can be reduced of processors) need not run in synchrony. As a result,by enriching the theory, thereby adding more constraints to be timing relations between processes can change from timesatisﬁed by any model. This may still leave different models, to time. For this and other reasons, Turing machines doof which only one, or a subset is intended. Such residual not provide good models for minds, as explained in .ambiguities are reduced (but never completely removed) by Moreover if sophisticated memory management systems arelinks between the theory and methods of observation and used the mappings between components of virtual machinesexperiment, and practical applications associated with the and physical parts of the system can also keep changing.theory. However those links do not deﬁne the concepts, If such complex changing relationships are useful in systemsbecause old methods of observation and old experimental we have designed and implemented we need to keep an openprocedures can be replaced by new ones, while the old mind as to whether evolution, which had several billion yearsconcepts endure. This loose, theory-mediated, connection head start on us, also discovered the power and usefulness of
such ﬂexibility. If so, some apparently important searches for can be causes’ by claiming that the objects, events andmind-brain correlations may turn out to be a waste of time. processes in virtual machines actually are physical objects, events and processes viewed in an abstract way. This is aCan virtual machines do things? ‘Mind-brain identity theory’, or a ‘virtual-physical machineIt is sometimes thought that the virtual machines in a computer identity theory’, sometimes presented as a ‘dual-aspect’ theory.do nothing: they are just ﬁgments of the imagination of A detailed rebuttal of this thesis requires showing (a) that itsoftware engineers and computer scientists. On this view, only is based on a bad theory of causation, (b) that the claim ofthe physicists and electronic engineers really know what exists identity being used here is either vacuous or false becauseand interacts in the machine. That derives from a widely held the actual relations between contents of virtual machines andtheory of causality, which assumes either that only physical contents of physical machines are quite unlike other casesevents can really be causes that produce effects, or, more where we talk of identity (e.g. because the virtual to physicalsubtly that if events are caused physically then they cannot mapping constantly changes even within the same machine)be caused by events in virtual machines. Some researchers (c) that the concepts used in describing virtual machinehave inferred that human mental events and processes, such phenomena (e.g. ‘software bug’, ‘failure to notice’, ‘preferringas weighing up alternatives, and taking decisions cannot have X to Y’, etc.) are too different from those of the physicalany consequences: they are mere epiphenomena. There is no sciences to be capable of referring to the same things. Morespace here for a full rebuttal of this view, but the key idea is obviously, those virtual machine concepts are not deﬁnablethat causation is not like some kind of ﬂuid or physical force using only concepts of the physical sciences.that ﬂows from one thing to another. Rather, the concept of X This is a very compressed defence of the claim that virtualcausing Y is a very subtle and complex concept that needs to machine events can have causal powers. There are morebe analysed in terms of whether Y would or would not have detailed discussions on the Birmingham Cogaff web site.2happened in various conditions if X did or did not occur, or ifX had or had not occurred, or if X does or does not occur in the Psychotherapy as virtual machine debuggingfuture. The truth or falsity of those ‘counterfactual conditional’ This design-based approach using the concept of a ‘virtualstatements depends in complex ways on the truth or falsity of machine’ can, in principle, justify techniques that deal with avarious laws of nature, which we attempt to express in our subset3 of human mental problems by manipulating virtualexplanatory theories. machines instead of manipulating brains using chemicals, However, in addition to their role in true and false electric shocks, etc. If a virtual machine is suitably designedstatements counterfactual conditionals may also play a role then it is possible to identify and in some cases repair, certainin instructions, intentions, or motives. Thus ‘Had he not ‘bugs’, or ‘dysfunctional’ processes, by interacting with thecooperated I would have gone ahead anyway’ may be not so running system. Such debugging is common in teaching.much a retrospective prediction as an expression of resolve. Doing that well requires deep, explanatory theories ofPlans, intentions and strategies involve causation as much as how normal mental virtual machines work. Some debuggingpredictions and explanations do. These too can play a causal techniques for minds have evolved through various kinds ofrole in virtual machines. social experimentation without such deep theories, and many of them work well. For instance, when a child gets the wrongReal causation in virtual machines results for long division we don’t call in a brain surgeon,Suppose one of your ﬁles disappears. A software engineer may but check whether he has perhaps learnt the wrong rule orconclude (after thorough investigation) that one of your actions misunderstood one of the mathematical concepts involved. Inactivated a bug in a running program, which led another part of such cases the lack of competence can be seen as a bug thatthe program to remove that ﬁle. These are events and processes can be ﬁxed by talking, drawing diagrams, and giving simplerin a virtual machine. This diagnosis could lead the programmer examples. It might be thought that this is unlike ﬁxing bugsto make a change in the operating system or ﬁle management on computers because in the latter case the process has to besystem that alters its future behaviour. Sometimes this can be stopped, the new program compiled and the program restarteddone by altering software rules without restarting the machine whereas humans learn and change without having to cease– rather like telling a person how to do better. functioning. But some programs are run using interpreters Typically, the causal connections discovered and altered by or incremental compilers, which allow changes to be madesoftware engineers do not require any action by electronic to the running program without stopping and restarting theengineers or physicists to alter the physical machine, and most process with a recompiled program. Several AI programmingof the people with expert knowledge about the hardware would languages have that kind of ﬂexibility. More sophisticatednot even understand how the bug had occurred and how it was software development tools can also plant mechanisms forﬁxed. interrogating and modifying running systems. Moreover, virtual machine components can constantly Of course, teaching someone how to improve his longchange their mapping onto hardware, so the change made 2 See this presentation on virtual machinesby the software engineer need not have any speciﬁc physical http://www.cs.bham.ac.uk/research/projects/cogaff/location when the software is running. talks/#inf and this discussion of free-will and causation http://www.cs.bham.ac.uk/research/projects/cogaff/The mind-brain identity defence misc/four-kinds-freewill.htmlSome philosophers defend the thesis that ‘only physical events 3 Don’t expect it to remove brain-tumours, for instance!
division, and giving counselling to enable a patient to values. Isn’t that what happens to humans? This point wasunderstand how he unintentionally causes family rows to made long ago in section 10.13 of . So motivation inescalate, differ in detail from the process of ﬁxing a typical artiﬁcial virtual machines, including self-generated motivation,bug in a computer program. For example the teaching and is not a problem in principle.therapy depend on kinds of self understanding that few The myth that intelligence requires emotionscomputer programs have at present. But in future it will In recent years, especially following publication of Damasio’snot be uncommon to ﬁnd virtual machines, that, instead  and Picard’s , much has been made of the alleged needof being forcibly altered by a user editing rules, instead to ensure that intelligent systems have emotions. I have arguedchange themselves as a result of being given advice about elsewhere that the arguments are fallacious for example in how to behave. That’s not even science ﬁction: it is not     and . Some of these claims are merelyhard to achieve. This comparison should undermine three poorly expressed versions of Hume’s unsurprising observationassumptions: (a) the belief that the only way to understand that without motives an intelligent system will have no reasonand ﬁx problems with minds is to work in a totally bottom to do anything: this is just a confusion between emotions andup way, namely understanding and modifying brains, (b) the motives.belief that nothing we know about computers is relevant to There is also a more subtle error. 26 years ago, understanding and repairing minds, and (c) the belief that argued that intelligent machines need mechanisms of kindsmachines can only be programmed to do things, not advised, that perform important functions and which in addition caninspired, or instructed or cajoled into doing them. sometimes generate emotional states and processes as sideMotivation in virtual machines effects of their operation, if they lack sufﬁcient processingSo, states, events, and processes occurring in virtual machines power to work out what needs to be done. This does not implycan have causal inﬂuences which alter both other things in that they need emotions,4 . just as the fact that some operatingthe virtual machine and also the physical behaviour of the systems need mechanisms that are sometimes capable ofsystem (e.g. physical events in memory, hard drives, internal generating ‘thrashing’ behaviour does not imply that operatinginterfaces, what is displayed on a screen, the sounds coming systems need thrashing behaviour. Desirable mechanisms canfrom speakers, or various attached motors, etc.). However, we sometimes have undesirable effects. There can be particularstill need to understand the variety of kinds of causation in sorts of situations where an ‘alarm system’ detects a putativevirtual machines. need to override, modulate, freeze, or abort some other In very simple computer models nothing happens until a process, and where because of shortage of information oruser gives a command, which can then trigger a cascade shortage of processing power, a rough and ready rule operates.of processes. In more sophisticated cases the machine is It would be better if the situation could be fully evaluateddesigned so that it initiates various activities from time and reasoned about (without any emotions), but if there isto time, e.g. checking for email, checking whether disks inadequate capacity to do that in the time available then itneed defragmenting, checking whether current scheduling may be better to take the risk of false alarms, especially ifparameters need to be revised in order to improve processing the alarm system has been well trained, either by evolution orperformance. It can also be designed so that events initiated individual learning, and does not often get things wrong.from outside trigger new internal processes. E.g. a user The need for therapy to undo bad learningattempting to access a ﬁle can trigger a sub-process checking However the danger in having such powerful subsystems thatwhether the user has the right to access that ﬁle. All of can change as a result of learning is that they may change inthese cases require the designers of the virtual machine to bad ways – as clearly happens in some humans. In some cases,anticipate kinds of things that might need appropriate checking it may be possible to undo bad changes by re-programming,or corrective actions to be performed. e.g. through discussion, advice, teaching, re-training, therapy, But there is no difﬁculty in building a machine that acquires etc. In other cases help may be available only when it isnew competences while it is running, and also new conditions too late. Of course there are also cases where such systemsfor exercising old competences. If the virtual machine go wrong because of physical damage, disease, malfunction,architecture allows new goal-generators to be acquired, new corruption, etc., which may or may not be reversible, e.g.strategies for evaluating and comparing goals, new values to chemical addictions.be employed in such processes, then after some time, themachine may have goals, preferences, intentions, etc. that were III. T HE VARIETY OF MENTAL VIRTUAL MACHINESnot given to it by anyone else, and which, as remarked in There is not just one kind of mind. Insect minds are different, can only be described as its own. It would be, to that from minds of birds and monkeys. All are different in manyextent, an autonomous machine, even if the processes by which ways from adult human minds. Human minds are differentsuch motives and motive-generators were acquired involved at different stages of development: a newborn infant, a nine-being inﬂuenced by things said and done by other people, for month old crawler, a two-year old toddler, a four year oldinstance teachers, heroic ﬁgures, and other role models. Indeed talker, a 50 year old professor of psychiatry. Apart froma machine might be designed speciﬁcally to derive motives,values, preferences, etc. partly on the basis of such inﬂuences, 4 As argued in this slide presentation: http://www.cs.bham.ac.uk/tempered by experimentation on the results of trying out such research/projects/cogaff/talks/#cafe04
differences resulting from development and learning there are real robots, but that what the researchers were trying to getdifferences that can be caused by genetic deﬁciencies, and by their systems to do fell very far short of what they impliedbrain malfunctions caused by disease or injury. What’s more in their predictions and promises would be achieved, becauseobvious is that the brains are different too. So we need to ﬁnd a they had not analysed the requirements adequately.general way of talking about different minds, different brains, Moreover, it could turn out that all of the currentlyand the different kinds of relationships that hold between (a) understood models of computation are inadequate, and entirelythe minds, i.e. the virtual machines, and (b) the brains, i.e. the new kinds of virtual machine are needed, including machinesphysical machines in which they are implemented. that grow their own architecture, as suggested in .Even virtual machines have architecturesWe are talking about a type of complex system with many IV. W HAT IS INFORMATION ?concurrently active parts that work together more or less There are many questions still to be answered about theharmoniously most of the time but can sometimes come concept of information used here. What is ‘information’? Whatinto conﬂict. These parts are organised in an information- is an information-user? What is involved in understandingprocessing architecture that maps onto brain mechanisms in something as expressing a meaning or referring to something?complex, indirect ways that are not well understood. So we Is there a distinction between things that merely manipulateshould ask questions like this if we wish to do deep science symbolic structures and things that manipulate them whilestudying a particular kind of mind: understanding them and while using the manipulation to derive• What sorts of component parts make up the architecture of new information? In how many different ways do organismsthis sort of mind? acquire, store, derive, combine, manipulate, transform and use• What are their functions? information? How many of these are, or could be, replicated• For each such component, what difference would it make if in non-biological machines? Is ‘information’ as important ait were modiﬁed or removed, or connected in a different way concept for science as ‘matter’ and ‘energy’, or is it justto other components? a term that is bandied about by undisciplined thinkers and• Which parts of the architecture are involved in various popularists? Can it be deﬁned? Is information something thatprocesses that are found to occur in the system as a whole? should be measurable as energy and mass are, or is it more• Which parts are connected to which others, and how do like structure, which needs to be described not measured (e.g.they interact? the structure of this sentence, the structure of a molecule, the• What kinds of information do the different parts acquire structure of an organism, the properties of a toroid)?and use, and how do they obtain the information? We currently have the paradoxical situation that philoso-• How is the information represented? (It could be phers who are good at conceptual analysis are badly informedrepresented differently in different subsystems). about and often have false prejudices about computing• Can the system extend the varieties of information contents systems, whereas many software engineers have a deep butthat it can make use of (extend its ontology), and extend the unarticulated understanding, which they use very well whenforms of representation that it uses? designing, implementing, testing, debugging, maintaining,• What is the total architecture in which they function, and virtual machines, but which they cannot articulate well becausehow is it made up of sub-architectures? they have not been taught to do philosophy.• How are the internal and external behaviours se- Information-theory is not about information!lected/controlled/modulated/coordinated? The word ‘information’ as used (after Shannon) in so-called• Can conﬂicts between subsystems arise, and if so how can ‘information theory’ does not refer to what is normallythey be detected, and how can they be resolved? meant by ‘information’, since Shannon’s information is a• What mechanisms in virtual machines make those processes purely syntactic property of something like a bit-string, orpossible, and how are they implemented in brains? other structure that might be transmitted from a sender to• In how many ways can the different virtual machines either a receiver using a mechanism with a ﬁxed repertoire ofindividually, or through their interactions go wrong, or produce possible messages. Having that sort of information does not,dysfunctional effects? for example, allow something to be true or false, or to• How did this virtual machine evolve, and what does it contradict something else. However, the more general concepthave in common with evolutionary precursors and with other of information, like ‘mass’, ‘energy’ and other deep conceptscontemporary animal species? used in scientiﬁc theories, is not explicitly deﬁnable. That is Answering these, and similar, questions requires a long term to say, there is no informative way of writing down an explicitinvestigation. One of the reasons why optimistic predictions deﬁnition of the form ‘X is Y’ if X is such a concept. All you’llregarding imminent successes of Artiﬁcial Intelligence (or end up with is something circular, or potentially circular, whenmore recently robotics) have repeatedly failed is that the you expand it, e.g. ‘information is meaning’, ‘information isphenomena we are trying to explain and to replicate, are far semantic content’, ‘information is aboutness’, ‘information ismore complex than anyone imagines. I.e. the main reason is what is expressed by something that refers’, ‘information is aNOT that the wrong programming languages, or the wrong difference that makes a difference’ (Bateson), and so on.models of computation were used (e.g. symbolic vs neural), But that does not mean either that the word is meaninglessor that the test implementations used simulations rather than or that we cannot say anything useful about it. The same is true
of ‘energy’. It is sometimes deﬁned in terms of ‘work’, but though U has never encountered S or anything with similarthat eventually leads in circles. So how do we (and physicists) information content, for instance when U encounters a newmanage to understand the word ‘energy’? sentence, diagram or picture for the ﬁrst time. Even before The answer was given above in Section I: we understand any user encounters S, it is potentially usable as an informationthe word ‘energy’ and related concepts, by understanding their bearer. Often, however, the potential cannot be realised withoutrole in a rich, deep, widely applicable theory (or collection U ﬁrst learning a new language, or notation, or a new theoryof theories) in which many things can be said about energy, within which the information has a place, and which providese.g. that in any bounded portion of the universe there is a substantive ontology extension for U, as discussed in .scalar (one-dimensional), discontinuously variable amount of This may be required for growth in self-knowledge too.it, that its totality is conserved, that it can be transmitted in A user with appropriate mechanisms has potential tovarious ways, that it can be stored in various forms, that it can derive inﬁnitely many distinct items of information from abe dissipated, that it ﬂows from objects of higher to objects small structure, e.g. inﬁnitely many theorems derivable fromof lower temperatures, that it can be used to produce forces Peano’s ﬁve axioms for arithmetic. Physically quite smallthat cause things to move or change their shape, etc. etc. As objects can therefore have inﬁnite information content, inscience progresses and we learn more about energy the concept combination with a reasoning mechanism, though limitationsbecomes deeper and more complex. The same is happening of the implementation (e.g. amount of memory available) maywith ‘information.’ constrain what is actually derivable. It follows that physicalAn implicitly deﬁned notion of ‘information’ structure does not constrain information content, unless a typeWe understand the word ‘information’ insofar as we use it in of user is speciﬁed.a rich, deep, and widely applicable theory (or collection of Information processing in virtual machinestheories) in which many things are said about information, Because possible operations on information are much moree.g. that it is not conserved (I can give you information complex and far more varied than operations on matter andwithout losing any), that instead of having a scalar measure energy, engineers discovered, as evolution had ‘discovered’of quantity, items of information, may form a partial ordering much earlier, that relatively unfettered information processingof containment (information I2 is contained in I1 if I2 is requires use of a virtual machine rather than a physicalderivable from I1), and can have a structure (e.g. there are machine. E.g. digital electronic calculators can perform farreplaceable parts of an item of information such that if those more varied tasks than mechanical calculators using cog-parts are replaced the information changes but not necessarily wheels.the structure), that two information items can share some It seems to be a basic law that increasing usable informationparts (e.g. ‘Fred hates Mary’ and ‘Mary hates Joe’), that it in a virtual machine by making implications explicit requirescan be transmitted by various means from one location or the physical implementation machine to use energy. Similarlyobject to another, that it can vary both discontinuously (e.g. (as suggested by Jackie Chappell), using greater informationadding an adjective or a parenthetical phrase to a sentence, content requires more energy to be used: e.g. in storage,like this) or continuously (e.g. visually obtained information sorting and processing information. So biological species ableabout a moving physical object), that it can be stored in to acquire and process vast amounts of information must bevarious forms, that it can inﬂuence processes of reasoning near the peak of a food pyramid, and therefore rare.and decision making, that it can be extracted from otherinformation, that it can be combined with other information Causal and correlational theories of meaning are falseto form new information, that it can be expressed in different It is often thought that learning to understand S as referringsyntactic forms, that it can be more or less precise, that it can to X, requires an empirical discovery that there is a causalexpress a question, an instruction, a putative matter of fact, relation between S and X, such as that occurrences of X alwaysand in the latter case it can be true or false, known by X, or often cause occurrences of S to come into existence, orunknown by Y, while Z is uncertain about it, etc. etc. such that the occurrence of X is shown empirically to be a reliable predictor of the occurrence of S. That this theoryInformation is relative to a user, or potential user is false is shown by the fact that you have no reason toWhereas energy and physical structures simply exist, whether believe that occurrences of the word ‘eruption’ are correlatedused or not, information in a physical or virtual structure with or reliable predictors of eruptions, and moreover you canS is only information for a type of user. Thus a structure understand the phrase ‘eruption that destroyed the earth 3000S refers to X or contains information about X for a user years ago’ even though it is impossible for any such correlationof S, U. The very same physical structure can contain or causal link to exist, since the earth was not destroyed then.different information for another user U’, or refer to something Further we can use concepts that refer to abstract entitiesdifferent for U’, as shown by ambiguous ﬁgures, and also whose existence is timeless, such as the number 99, or thewritten or spoken languages or notations that some people shortest proof that 2 has no rational square root. So empiricalunderstand and others do not. This does not make information correlations and causal inﬂuences are impossible in thoseinherently subjective any more than mountains are subjective cases.because different people are capable of climbing differentmountains. (Indexicality is a special case, discussed below.) Information content determined partly by contextThe information in S can be potentially usable by U even It is sometimes thought that artiﬁcial minds would never
be able to grasp context-sensitive information. For example, parts cannot access. When we ask ‘Did X know that P?’ itan information-bearing structure S can express different is not clear whether the answer must be ‘yes’ in cases whereinformation, X, X’, X”, for the same user U in different some part of X made use of the information that P. E.g. humancontexts, e.g. because S includes an explicit indexical element posture control mechanisms use changes in optical ﬂow. Does(e.g. ‘this’, ‘here’, ‘you’, ‘now’, or non-local variables in that mean that when you are walking around you know abouta computer program). Indexicality can make information optical ﬂow even though you don’t know that you know it?incommunicable in the sense that the precise information Your immune system and your digestive system and variouscontent of one user cannot be transferred to another. (Frege, metabolic processes use information and take decisions offor example, showed that one user’s use of the word “I” has many kinds. Does that mean that you have the information, thata sense that another person is incapable of expressing.) A you know about the information, that you use the information?corollary is that information acquired by U at one time may Some people might say yes others no, and some may saynot be fully interpretable by U at another time, because the that it depends on whether you know that you are using thecontext has changed, e.g. childhood ‘memories’. In  it was information.argued that such indexicality accounts for the ‘ineffability’ of Likewise there are different parts of our brains that evolvedqualia. However this does not usually prevent the intended at different times that use different kinds of informationfunction of communication from being achieved. The goal of (even information obtained via the same route, e.g. thecommunication is not to replicate the sender’s mental state, or retina or ear-drum, or haptic feedback). Some of them areinformation content, in the receiver, but to give the receiver evolutionarily old parts, shared with other species (e.g. postureinformation that is adequate for some purpose. control mechanisms), some newer and possibly some unique Many structures in perceptual systems change what to humans, (e.g. human face recognition mechanisms, andinformation they represent as the context changes. Even if mechanisms that can learn to read music and other notations).what is on your retina is unchanged after you rapidly turn A deep feature of at least the human architecture, andyour head 90 degrees in a room, the visual information will perhaps some others also, is that sensors and effectorsbe taken to be about a different wall – with the same wallpaper providing interfaces to the environment can be shared betweenas the ﬁrst wall. Many examples can be found in . different subsystems, as described in . E.g. your vision Sometimes U takes S to express different meanings in mechanisms can be shared between: posture control, visualdifferent contexts because S includes a component whose servoing of manipulation actions, mechanisms involved insemantic role is to express a higher order function which reading instructions, and affective mechanisms that appreciategenerates semantic content from the context, e.g. ‘He ran aesthetic qualities of what is seen. Your walking mechanismsafter the smallest pony’. Which pony is the smallest pony can be shared between a subsystem concerned with movingcan change as new ponies arrive or depart. More subtly what to the door and also a subsystem concerned with socialcounts as a tall, big, heavy, or thin something or other can vary communication, e.g. ﬂirting by walking suggestively. We canaccording to the range of heights, sizes, weights, thicknesses describe such architectures as using ‘multi-window’ perceptionof examples in the current environment. and ‘multi-window’ action, whereas current artiﬁcial systems There are many more examples in natural language that lead mostly use only ‘peephole’ perception and ‘peephole’ action,to incorrect diagnosis of words as vague or ambiguous, when where input and output streams from each sensor or two eachthey actually express precise higher order functions, applied effector are go along channels of restricted functionality.to sometimes implicit arguments, e.g. ‘big’, ‘tall’, ‘efﬁcient’, Sometimes the sharing is concurrent and sometimes‘heap’, or ‘better’ (discussed in ).5 sequential. Conﬂict resolution mechanisms may be requiredInformation content shared between users when concurrent sharing is impossible. Some of theDespite the above, it is sometimes possible for X to mean the mechanisms that detect and resolve conﬂicts may besame thing to different users U and U’, and it is also possible inaccessible to self-monitoring (discussed later), so that anfor two users who never use the same information bearers individual may be unaware of important decisions being taken.(e.g. they talk different languages) to acquire and use the same Much philosophical, psychological, and social theorisinginformation. This is why relativistic theories of truth are false: misguidedly treats humans as unitary information users,although I can believe that my house has burned down while including Dennett’s intentional stance and what Newell refersmy neighbour does not, one of us must be mistaken: it cannot to as ‘the Knowledge level’.be true for me that my house has burned down but not truefor my neighbour. Truth is not ‘for’ anyone. Meaning depends Just the beginning of an analysis of ‘information’on the user. Truth does not. The analysis of the concept of ‘information’ presented here amounts to no more than a small fragment of the V. I NFORMATION - USING SUBSYSTEMS full theory of types of states, events, processes, functions, An information-user can have parts that are information mechanisms and architectures that are possible in (virtualusers. A part can have and use some information that other and physical) information-processing machines. I doubt that anyone has produced a clear, complete and deﬁnitive 5 This idea is developed in the context of Grice’s theory of communication, list of facts about information that constitute an implicitwith implications for the evolution of language, here: http://www.cs.bham.ac.uk/research/projects/cosy/papers/#dp0605 Spa- deﬁnition of how we (the current scientiﬁc community well-tial prepositions as higher order functions. educated in mathematics, logic, psychology, neuroscience,
biology, computer science, linguistics, social science, artiﬁcial brains, add only more of the same. However, in humans andintelligence, physics, cosmology, ...) currently understand the many other animals there additional kinds of information andword ‘information’. information processing. E.g. there’s a great deal still to be said about the Brains are needed for more than movementmolecular information processing involved in development of Most organisms manage without brains. Why not all?an individual from a fertilised egg or seed, and in the huge One function is resolving conﬂicts between different partsvariety of metabolic processes including intrusion detection, responsible for decisions that could be incompatible, e.g.damage repair, transport of materials and energy, and control decisions to move one way to get to food or to move anotherby hormones and neurotransmitters. It may be that we shall way to avoid a sensed predator. This requires coordination,one day ﬁnd that far more of the brain’s information processing possibly based on dynamically changing priorities. A differentis chemical than anyone dreams now is possible. kind of requirement, discussed later, is doing more complex A more complete theory would provide a more complete processing of information, e.g. in order to acquire a betterimplicit deﬁnition of the concept ‘information’ required understanding of the environment, or to acquire somethingfor understanding natural and artiﬁcial systems (including like a terrain map, or to plan extended sequences of actions.far more sophisticated future artiﬁcial systems). A hundred In organisms with many complex parts performing differentyears from now the theory may be very much more deep functions it may also be necessary to coordinate internaland complex, especially as information processing machines changes, for example all the changes involved in reachingproduced by evolution still seem to be orders of magnitude puberty in humans, or during pregnancy, where many internalmore complex than any that we so far understand.6 changes and external behaviours need to be coordinated. Lewis Wolpert wrote, in an Observer book review, March VI. B IOLOGICAL INFORMATION PROCESSING 24, 2002: ‘First the only function of the brain from an All living things, including plants and single-celled evolutionary point of view is to control movement and soorganisms, process information insofar as they use sensors to interaction with the environment. That is why plants do notdetect states of themselves or the environment and use that have brains.’7 This often quoted, but grossly misleading, claiminformation either immediately or after further processing to reﬂects a widespread current focus on research that modelsselect from a behavioural repertoire. The behaviour may be brains as sensory-motor control mechanisms, e.g. . Evenexternally visible physical behaviour or internal processes. if the original function of brains was to control movement, While using information an organism normally also uses up much of what human brains do has nothing to do with controlstored energy (usually chemical energy), so that it also needs of movement – for example explaining what is observed,to use information to acquire more energy. predicting future events, answering a question, and doing There are huge variations both between the kinds of mathematics or philosophy.information contents used by different organisms and between So we see that information in organisms may be implicitdifferent ways in which information is acquired, stored, and transient, implicit and enduring, explicit and capable ofmanipulated and used by organisms. The vast majority of multiple uses, used only locally, used in controlling spatiallyorganisms use only two kinds of information: (a) genetic distributed information processing, and may vary in kind ofinformation acquired during evolution, used in replication, abstraction, in the ontology used, in the form of representationphysical development and maintenance of the individual, and used and in the kinds of manipulation that are available.(b) transiently available information used by online control Varieties of explicit informationsystems. I call the latter ‘implicit’ information. There may Only special conditions bring about evolution of mechanismsalso be some less transient implicit information produced by that create, store and use explicit, that is enduring, re-usable,gradually adjusted adaptive mechanisms. information structures that can be used in different ways Since the vast majority of species are micro-organisms (e.g. forming generalisations, making predictions, buildingthe vast majority of information-using organisms can use terrain maps, forming motives, making plans, rememberingonly implicitly represented information; that is to say they what happened when and where, communicating to otheruse only information that is available during the transient individuals, etc.) As explained in , this requires newstates of activation produced by information being acquired architectures involving mechanisms that are not so directlyand used, and the information is represented only transiently involved in the sensorimotor relationships and their control.8in activation states of sensors, motors and intervening There may have been intermediate stages of evolutionmechanisms, and also in parameters or weights modiﬁed by in which non-transient information was stored only in theadaptive feedback mechanisms. environment, e.g. in chemical trails and land-marks used as The short term transient implicit information in patterns of sources of information about the presence of food or predators,activation and the longer term implicit information in gradually or about the routes followed by con-speciﬁcs (pheromonechanging adaptive mechanisms together sufﬁce for most living trails), or about location relative to a nest or an enduring sourcethings – most of which do not have brains! In some animals 7 http://observer.guardian.co.uk/science/story/0, 6 For a draft attempt to answer the question ‘What is infor- ,673268,00.htmlmation?’ see http://www.cs.bham.ac.uk/research/projects/ 8 See ‘Sensorimotor vs objective contingencies’ http://www.cs.cogaff/misc/whats-information.html bham.ac.uk/research/projects/cosy/papers/#dp0603
of food. (Some insects can use land-marks so this capability individual B believes is the case, but the functional rolesprobably evolved a long time ago.) This is possible only for of those two forms of representation will be different.9 Inanimals in an environment with stable structures, unlike some particular, the former will be ‘referentially transparent’ and themarine environments. latter ‘referentially opaque’, and the conditions for truth of the Mechanisms that evolved to use external enduring two beliefs are quite different – something young children doinformation may have been precursors to mechanisms not learn in the ﬁrst few years.using internally stored explicit information. There probably The ability to think and reason about the mental contentswere many different evolutionary transitions, adding extra of another has much in common with the ability to think andfunctionality. In previous papers (e.g. ) colleagues and reason about one’s own mental contents as far as mechanismsI have emphasised three main categories of competence and formalisms are concerned, though obviously differentrequiring different sorts of architectural components, namely sensor mechanisms are involved, i.e. external and internal.reactive, deliberative and meta-management capabilities, but However the evolutionary beneﬁts are very different. It isthere are many intermediate cases and different sorts of not clear which evolved ﬁrst. Probably both types of meta-combinations of cases that need to be understood – not only semantic competence co-evolved, each helping to enrich thefor understanding how things work when everything functions other.normally, but also in order to understand the many ways things Incidentally, nothing said here implies that any organismcan go wrong, including both the consequences of physical has full self-knowledge. On the contrary, self-knowledgemalfunctions and also the consequences of dysfunctional will inevitably be associated with a bottleneck with limitedprocessing of information in virtual machines. capacity – a point that has relevance for psycho-analysis. Compare .Deliberative mechanismsA very small subset of organisms (and some machines) have Perhaps a more widely-shared ability to formulate internala ‘deliberative’ information-processing capability insofar as questions or goals was an evolutionary precursor to meta-they can construct a set of information structures of varying semantic competence, since formulating a yes-no questioncomplexity (e.g. plans, predictions, theories), then compare (‘Will it rain today?’) requires an ability to representtheir merits and de-merits in relation to some goal, and propositions that are capable of being true or false withoutproduce new information structures describing those tradeoffs, a commitment to their truth value. The same is true of havingthen select one of the alternatives and make use of it. That desires and intentions. So the ability to represent ‘X thinksability to represent and reason about hypotheticals was one I will eat him’ and to reason about the consequences of X’sof the ﬁrst aspects of human processing modelled in AI thinking that (e.g. X will run away) may have arisen as asubsystems, but it was probably one of the last to evolve, modiﬁcation of the ability to represent one’s own goals (‘Iand only a very tiny subset of organisms have it in its richest will eat X’) that have not been achieved, plans that have notform, described in . yet been carried out, tentative predictions that have not yet It is now fashionable to contrast that early ‘symbolic’ AI been tested, and questions that have not yet been answered.work with biologically inspired AI work. But that is just silly, Old and new in the same architecturesince all of the work was biologically inspired insofar as it was It is not always remembered that besides having suchan attempt to model processes that occur in human beings. A sophisticated (and possibly unique) capabilities for explicithuman making a plan or proving a theorem is just as much a manipulation of information, humans share many informationbiological organism as a human running, jumping or catching processing capabilities with other species that lack thea ball, even though more species share the capabilities required distinctively human capabilities. For instance many animalsfor the latter activities. Biology is much richer than some have excellent vision, excellent motor control, abilities toresearchers seem to realise. cooperate on certain tasks, hunting capabilities, the abilityMeta-semantic competences to avoid predators, nest-building capabilities, as well asAnother relatively rare biological information-processing the information-processing involved in control of bodilycapability involves the ability to refer to, reason about, or functions. The notion that somehow human cognition, orcare about things that themselves contain, or use information. human conscious processes can be studied and replicatedFor instance when you discover what someone thinks, and without taking any account of how these more general animalwhen you worry about someone’s motives you are using mechanisms work and how they interact with the distinctivelysuch meta-semantic competence. Humans can also apply this human mechanisms, may lead both to a failure to understandmeta-semantic competence to themselves (though probably humans and other animals and also to designs for robots thatnot at birth: the architecture needs time to develop). Having don’t perform as required.this ability requires a more complex architecture than merely Some of the ideas developed with colleagues in Birminghambeing able to refer to things without semantic content (e.g. about the different components in such a multi-functionalphysical objects and processes). That’s partly because having architecture are reported in the paper by Brigitte Lorenz andmeta-semantic content involves representing things that are Etienne Barnard presented at this conference and will not betreated as true in only a hypothetical, encapsulated, way. The 9 J. Barnden’s ATT-META project has developed a way of making thatvery same form of representation may be used for what an distinction which is also related to the ability to think metaphorically. http:individual A believes is the case and for what A thinks another //www.cs.bham.ac.uk/∼jab/ATT-Meta/
developed here. From this standpoint it is almost always a to run, if an interpreter is available. Moreover if it weremistake to ask questions like ‘How do humans do X?’ Instead necessary, then substantive ontology extension during learningwe should ask ‘How do different subsystems in humans do and development would be impossible.X?’ (e.g. X = control actions, interpret visual input, learn, store Nobody is yet in a position to say what the prelinguisticinformation, react to interruptions, generate motives, resolve languages do and do not share with human communicativeconﬂicts, etc.) And we should expect different answers not languages. In particular, it may be the case that theonly for different subsystems, as in Trehub’s , but also main qualitative competences required for human languagefor humans at different stages of development, in different use already exist in these pre-linguistic competences incultures, with and without brain damage, etc. young children and other animals, and that the subsequent evolutionary developments related to human language were mainly concerned (a) with developing means of VII. P RE - LINGUISTIC COMPETENCES generating adequately articulated external behaviours to It is often assumed that there is a massive discontinuity communicate their structures, (b) improving perception ofbetween human linguistic competence and other competences such behaviours, and (c) extending internal mechanisms (e.g.(e.g. as argued by Chomsky in ), though there are many short term memories) that are equally useful for sophisticatedwho hotly dispute this (e.g. Jablonka and Lamb ). But internal information processing (such as planning) and forpeople on both sides of the dispute make assumptions about communication with others.human language that may need to be challenged if we are If so, information (unconsciously) acquired, used and storedto understand what human minds are, and if we wish to for later use in early childhood may have much richerproduce working human-like artiﬁcial minds. In particular it structures and deeper semantic content than has previouslyis often assumed that the essential function of language is been thought possible. Whether it includes the kind of contentcommunication. But, as argued in , and more recently required to support psychoanalytic theories remains open.in  many of the features that make such communication Non-auditory communicationpossible (e.g. the ability to use varieties of information with It is easier to achieve a large collection of perceptuallyrich and varied structures and with compositional semantics, discriminable signals by using independently movable ﬁngers,and non-communicative abilities to check whether some state hands, mouth, eyes, head than to do it all by modulatingof affairs matches a description, to notice a gap in information, an acoustic signal – especially as that would interfere withexpressed in a question) are also requirements for forms breathing. So, if some animals already had hands for whichof information that are involved in perception, planning, they were learning to produce (and perceive, during controlledexpressing questions, formulating goals, predicting, explaining execution) large numbers of distinct manipulative competencesand reasoning. concerned with obtaining and eating food, grooming, climbing, From that viewpoint, communication between individuals making nests, ﬁghting, threatening, etc., then perhaps thein a public medium was a secondary function of language ﬁrst steps to communicative competence used structuredwhich evolved only after a more basic kind of competence movements, as in that involved producing and perceivingevolved – the ability to use an internal, non-communicative, structured sign language, rather than vocal language.language in mental states and processes such as perceiving, There is much suggestive evidence, including the fact thatthinking, supposing, intending, desiring, planning, predicting, humans ﬁnd it almost impossible not to move hands, eyes,remembering, generalising, wanting information, etc. and in facial muscles, etc. when talking, even when talking on theexecuting intentions or plans. phone to someone out of sight. More compelling is the ease It is clear that human children who cannot yet talk, and with which deaf babies are able to learn a sign language, andmany animals that do not use an external language can the reported fact that some children with Down syndrome seemperceive, learn, think, anticipate, have goals, threaten, be to learn to communicate more easily if sign language is used.puzzled, play games, carry out intentions and learn and use Most compelling is the case of the Nicaraguan deaf childrenfacts about causation. I know of no model of how any of that who invented their own highly sophisticated sign language,can be done without rich information processing capabilities leaving their teacher far behind. See .10of kinds that require the use of internal languages with An implication of this is that the human ability to developcompositional semantics. But I know of no detailed model linguistic competence does not depend on learning a languageof what those ‘prelinguistic’ languages are, how they evolved, that is already in use by others. Language learning thenhow they develop, how they work, etc. appears to be a process of collaborative, creative problem- Fodor, in , postulated a ‘language of thought’ (LOT) solving, which may be constrained by the prior existence of awhich was supposed to be innate, available from birth and social language, but need not be.capable of expressing everything that ever needs to be said In  Arbib proposed that action recognition and imitationby any human being, but he left most questions about how was a precursor to the evolution of language, but the argumentsit worked unanswered and was not concerned with non-human animals. Moreover he supposed that external languages 10 http://www.indiana.edu/∼langacq/E105/Nicaragua.are translated into the LOT, whereas there is no reason html ‘A Linguistic Big Bang’, by Lawrence Osborne The New York Times October 24, 1999. A ﬁve minute video including examples of the inventedto believe such translation is necessary, just as compiling language is available here http://www.pbs.org/wgbh/evolution/to machine code is not necessary for computer programs library/07/2/l 072 04.html
are somewhat different, and do not share our hypothesis in context demands, in collaborative problem solving rather that the existence of a rich non-communicative language than in a process of information transfer. (This is related topreceded the evolution of communicative language, though Grice’s maxims of communication, and to current concernsthe commentary by Bridgeman  makes a similar point, with ‘situatedness’ in language understanding, perceptionemphasising pre-linguistic planning capabilities. and action. http://www.cs.bham.ac.uk/research/ Once something like this form of structured communication projects/cosy/papers/#dp0605)had developed, enabling information that was previously If the contents of linguistic communication betweenonly represented in internal languages to be shared between individuals make such use of the context then it may beindividuals via a public language, the enormous advantages equally true for thought processes within an individual. In thatof having such a communicative competence, both in solving case, much of the information content of your mind is outsideproblems that required collaboration and in accelerating you. This may be one of the ideas people who emphasisecultural transmission, might have led to an unusually rapid the importance of ‘embodiment’ are getting at. However,process of selection for changes that enhanced the competence everything said here could apply to minds in individualsin various ways, e.g. developing brain regions specialised at with simulated bodies located in simulated environments. Thestorage of ever larger vocabularies and more complex rules for important thing is what sorts of structures, processes andconstruction and interpretation of action sequences. causal interactions occur within the individual’s information Another development might have been evolution of brain processing architecture and how they are related to itsmechanisms to support construction and comparison of environment. Whether that environment is physical or anothermore complex symbolic structures, e.g. more deeply nested virtual machine environment makes little or no difference tostructures, and structures with more sub-structures. the kind of mind discussed here.The nature of linguistic communication From sign language to sound languageIf linguistic communication evolved from collaborative non- Use of sign language has many problems. As has often beenlinguistic activities, controlled by sophisticated internal noted, it can work only when the sender’s hands are free,languages, this must change our view of human language. and when the receiver can see the sender, thereby ruling outIt is often thought that linguistic communication involves a communication in many situations where visual attention mustprocess whereby an information structure in one individual be used for another task, or where there are obstructionsgets encoded in some behaviour which is perceived and to vision, or no light is available. So perhaps our ancestorsdecoded by another individual who then constructs the same started replacing hand signals with sounds where possible.information internally. However, in collaborative, creative, This could then lead in the usual way to a succession ofproblem solving the shared physical and task contexts provide evolutionary changes in which the vocal mechanisms, auditorymuch of the information, and need not be communicated. All mechanisms, and the neural control systems evolved alongthat is needed is whatever sufﬁces to produce desired results, with physiological changes to the vocal tract, etc. But thewithout copying an information structure from one brain to ability to learn sign languages remains intact, even if not usedanother. by most people. If a mother hands a child a piece of fruit, the child may If all this is correct then it may have deep implicationsthereby be triggered to form the goal of eating it, without the for clinical developmental psychology as well as formother having had the goal of eating it. Moreover, the mother understanding precisely what the effects of various kinds ofdoes not need to anticipate the precise movements of bringing brain damage are.the food to the mouth nor the precise chewing and swallowingmovements produced by the child. Those details can be leftto the child. More generally, context allows communication to VIII. A RCHITECTURES THAT GROW THEMSELVESbe schematic rather than concrete and detailed. This is obvious when A asks ‘Where are my keys?’ and Much research in AI and robotics is concerned with whatB replies ‘Look behind you?’. B may not know exactly sort of architecture an intelligent system needs, and variouswhere the keys are except that they are on the table behind rival architectures are proposed. If the architecture needs toA. A turns round and sees the keys, getting the precise grow itself, some of this research effort may have been wasted.information he wanted, as a result of B’s communication, even It was argued in  that the architecture should not be regardedthough B did not have that information. Both have achieved as genetically determined, but as grown using both innatetheir communicative goals, but not by transmitting a speciﬁc and acquired meta-competences inﬂuenced by the physicalinformation structure from B to A. and cultural environment, as depicted in Figure 1. So there So, many kinds of linguistic and non-linguistic commu- may be considerable differences in how adult minds work,nication involve communication of a partial or schematic as a result of a different developmental trajectories causedstructure, leaving gaps to be ﬁlled by the receiver using by different environments. This might affect kinds of self-information that may or may not be available to the monitoring, kinds of self control, kinds of learning ability,sender. In that case, many linguistic expressions may be and so on. Another implication is that attempts to quantify thebest thought of as having a higher-order semantics, to inﬂuence of genes and environment in terms of percentages isbe applied to more speciﬁc non-linguistic information as completely pointless.
presentations are in the Birmingham CogAff and CoSy project web sites. Jackie Chappell and Gerhard Pratl made useful comments on an early draft.Fig. 1. This indicates some of the many and varied relationships betweenthe genome and behaviours, produced at different stages of development afterlayers of competences and meta-competences have been built up. (From ) IX. C ONCLUSION : MACHINES IN GHOSTS Although I was asked to report on developments in AI thatmight be relevant to psycho-analysis I have instead focusedon some very general features of requirements for AI modelsrather than speciﬁc AI models, because I think that despite allthe advances achieved in AI, the models are still very primitivecompared with human minds and have a very long way togo before they can be directly relevant to working therapistsdealing with real humans. Nevertheless, I hope it is now clear why a human-likemachine, a human-like intelligent ghost, and indeed a human-like human must contain very speciﬁc kinds of informationprocessing virtual machines. So our ability to understandhuman minds and ways they can go wrong, including acquiringfalse, beliefs, inappropriate motives, inadequate strategies,tendencies to be over-emotional, and worse, must be informedby deep knowledge about information-processing systems ofthe appropriate kinds. The complexity of the machine, and the fact that self-monitoring mechanisms within it provide only a very limitedsubset of information about what is going on cause manyindividuals to start wondering what is really going on in them.Ignorance about information-processing mechanisms can makethings seem so mysterious that some people invoke a specialkind of stuff that is quite unlike physical matter. From there itis a short step to immortality, souls, etc., and the conceptualconfusions discussed by Ryle in . But if we don’t go thatfar we may still come up with deep new ways of thinking aboutphenomena that were previously thought resistant to computermodelling. This could, among other things, lead to a revolutionin psychoanalysis. ACKNOWLEDGEMENT Thanks to Dean Petters, Jackie Chappell, Matthias Scheutz,Margaret Boden, members of the EU-Funded Cosy Project,and other collaborators and discussants over the years. Someof the ideas here were inspired by John McCarthy and MarvinMinsky, as well as Immanuel Kant, . Related papers and
R EFERENCES  N. Chomsky, Aspects of the theory of syntax. Cambridge, Mass: MIT Press,, 1965.  D. C. Dennett, Brainstorms: Philosophical Essays on Mind and  E. Jablonka and M. J. Lamb, Evolution in Four Dimensions: Genetic, Psychology. Cambridge, MA: MIT Press, 1978. Epigenetic, Behavioral, and Symbolic Variation in the History of Life.  A. Sloman, The Computer Revolution in Philosophy. Has- Cambridge MA: MIT Press, 2005. socks, Sussex: Harvester Press (and Humanities Press), 1978,  A. Sloman, “The primacy of non-communicative language,” in The http://www.cs.bham.ac.uk/research/cogaff/crp. analysis of Meaning: Informatics 5 Proceedings ASLIB/BCS Conference,  J. Chappell and A. Sloman, “Natural and artiﬁcial Oxford, March 1979, M. MacCafferty and K. Gray, Eds. London: Aslib, meta-conﬁgured altricial information-processing systems,” 1979, pp. 1–15, http://www.cs.bham.ac.uk/research/projects/cogaff/81- International Journal of Unconventional Computing, 2007, 95.html. http://www.cs.bham.ac.uk/research/projects/cosy/papers/#tr0609.  A. Sloman and J. Chappell, “Computational Cognitive Epigenetics  A. Sloman and J. Chappell, “The Altricial-Precocial Spectrum for (Commentary on ),” Behavioral and Brain Sciences, 2007. Robots,” in Proceedings IJCAI’05. Edinburgh: IJCAI, 2005, pp. 1187–  J. Fodor, The Language of Thought. Cambridge: Harvard University 1192, http://www.cs.bham.ac.uk/research/cogaff/05.html#200502. Press, 1975.  S. Harnad, “The Symbol Grounding Problem,” Physica D, vol. 42, pp.  L. Osborne, “‘A Linguistic Big Bang’,” The New York Times, 24 October 335–346, 1990. 1999, http://www.indiana.edu/ langacq/E105/Nicaragua.html. A ﬁve  I. Kant, Critique of Pure Reason. London: Macmillan, 1781, translated minute video including examples of the invented language is available (1929) by Norman Kemp Smith. at http://www.pbs.org/wgbh/evolution/library/07/2/l 072 04.html.  A. Sloman, “The irrelevance of Turing machines to AI,” in Computa-  M. A. Arbib, “From monkey-like action recognition to human language: tionalism: New Directions, M. Scheutz, Ed. Cambridge, MA: MIT An evolutionary framework for neurolinguistics,” Behavioral and Brain Press, 2002, pp. 87–127, http://www.cs.bham.ac.uk/research/cogaff/00- Sciences, vol. 28, no. 2, pp. 105–124, 2005. 02.html#77.  B. Bridgeman, “Action planning supplements mirror systems in language  A. Damasio, Descartes’ Error, Emotion Reason and the Human Brain. evolution,” Behavioral and Brain Sciences, vol. 28, no. 2, pp. 129–130., New York: Grosset/Putnam Books, 1994. 2005.  R. Picard, Affective Computing. Cambridge, Mass, London, England:  G. Ryle, The Concept of Mind. London: Hutchinson, 1949. MIT Press, 1997. A. Sloman and M. Croucher, “Why robots will have emotions,” in Proc 7th Int. Joint Conference on AI. Vancouver: IJCAI, 1981, pp. 197–202. A. Sloman, “Towards a grammar of emotions,” New Universities Quarterly, vol. 36, no. 3, pp. 230–238, 1982, http://www.cs.bham.ac.uk/research/cogaff/96-99.html#47. A. Sloman, “Real time multiple motive-expert systems,” in Proceedings Expert Systems 85, M. Merry, Ed. Cambridge University Press, 1985, pp. 213–224. A. Sloman, “Damasio, Descartes, alarms and meta-management,” in Proceedings International Conference on Systems, Man, and Cybernetics (SMC98), San Diego. IEEE, 1998, pp. 2652–7. A. Sloman, R. Chrisley, and M. Scheutz, “The architectural basis of affective states and processes,” in Who Needs Emotions?: The Brain Meets the Robot, M. Arbib and J.-M. Fellous, Eds. Oxford, New York: Oxford University Press, 2005, pp. 203–244, http://www.cs.bham.ac.uk/research/cogaff/03.html#200305. A. Sloman, “Do machines, natural or artiﬁcial, really need emotions?” 2004, Invited talk: http://www.cs.bham.ac.uk/research/projects/cogaff/talks/#cafe04. A. Sloman, “‘Ontology extension’ in evolution and in development, in animals and machines,” University Aaron Sloman Aaron Sloman studied for a BSc in mathematics and physics of Birmingham, UK, 2006, PDF presentation: at Cape Town in 1956, then, after ﬂirting for a while with mathematical http://www.cs.bham.ac.uk/research/projects/cosy/papers/#pr0604. logic was seduced by philosophy, and completed a D.Phil at Oxford in 1962, A. Sloman and R. Chrisley, “Virtual machines and consciousness,” defending Kant’s philosophy of mathematics. He later learnt that AI provided Journal of Consciousness Studies, vol. 10, no. 4-5, pp. 113–172, 2003. the best tools for doing philosophy and since about 1970 has been working on A. Berthoz, The Brain’s sense of movement, ser. Perspectives in interdisciplinary problems concerned with understanding the space of possible Cognitive Science. London, UK: Harvard University Press, 2000. designs for behaving systems including humans, other animals and possible A. Sloman, “How to derive “better” from “is”,” future robots. In recent years this has particularly involved collaborating with American Phil. Quarterly, vol. 6, pp. 43–52, Jan 1969, a biologist studying cognition in birds and roboticists designing robots able to http://www.cs.bham.ac.uk/research/cogaff/sloman.better.html. perceive and manipulate 3-D objects in the environment. The various strands A. Sloman, “The mind as a control system,” in Philosophy and the of his work are summarised in http://www.cs.bham.ac.uk/∼axs/ Cognitive Sciences, C. Hookway and D. Peterson, Eds. Cambridge, my-doings.html He is a Fellow of AAAI, ECCAI and SSAISB, and was UK: Cambridge University Press, 1993, pp. 69–110. given an honorary DSc by Sussex University in 2006. M. Lungarella and O. Sporns, “Mapping information ﬂow in sensorimotor networks,” PLoS Computational Biolology, vol. 2, no. 10:e144, 2006, dOI: 10.1371/journal.pcbi.0020144. A. Sloman, “Sensorimotor vs objective contingencies,” University of Bimringham, School of Computer Science, Tech. Rep. COSY-DP-0603, 2006, http://www.cs.bham.ac.uk/research/projects/cosy/papers/#dp0603. A. Sloman, “Architecture-based conceptions of mind,” in In the Scope of Logic, Methodology, and Philosophy of Science (Vol II). Dordrecht: Kluwer, 2002, pp. 403–427, (Synthese Library Vol. 316). A. Sloman, “Requirements for a Fully Deliberative Architecture,” University of Bimringham, School of Computer Science, Tech. Rep. COSY-DP-0604, 2006, http://www.cs.bham.ac.uk/research/projects/cosy/papers/#dp0604. M. Minsky, “Matter Mind and Models,” in Semantic Information Processing, M. Minsky, Ed. Cambridge, Mass,: MIT Press,, 1968. A. Trehub, The Cognitive Brain. Cambridge, MA: MIT Press, 1991, Online http://www.people.umass.edu/trehub/.